Course Overview
NVIDIA GPU Stack & CUDA takes you under the hood of the compute layer powering every NVIDIA robotics product. You will write CUDA kernels, optimise neural network inference with TensorRT, serve models at scale with Triton Inference Server, build video analytics pipelines with DeepStream, and master GPU memory management and profiling tools. This course is for engineers who need to squeeze maximum performance out of NVIDIA hardware.
Key Highlights
✓
Write and debug custom CUDA C++ kernels for robot data processing
✓
Optimise neural networks with TensorRT: FP16 and INT8 quantisation
✓
Deploy model serving at scale with Triton Inference Server
✓
Build real-time video analytics pipelines with DeepStream SDK
✓
Profile GPU utilisation with Nsight Systems and Nsight Compute
✓
Manage GPU memory: unified memory, streams, and async operations
Course Syllabus
Learning Outcomes
01
Write efficient CUDA kernels for custom robot data processing
02
Optimise and deploy neural networks with TensorRT
03
Serve AI models at production scale with Triton
04
Build real-time video analytics pipelines with DeepStream
05
Profile and debug GPU performance bottlenecks
06
Design GPU memory layouts for maximum throughput
Who Is This For?
👤
Systems Engineers optimising robot AI performance
👤
ML Engineers deploying models to edge GPU devices
👤
Platform Teams building AI inference infrastructure
👤
Senior Developers who want GPU programming expertise
Prerequisites
- ☑ Proficiency in C++ and Python
- ☑ Understanding of neural network architectures
- ☑ Basic Linux sysadmin skills
- ☑ Experience with any deep learning framework (PyTorch or TensorFlow)